Build generic benchmark runner for nForma skills

[jobordu]

Problem

nf:solve has a CI benchmark gate that catches regressions before they ship. The same risk exists for quick, quick --full, new-milestone, fix-tests, and debug — but no equivalent coverage exists today.

User Story

As a maintainer, I want every PR to run smoke benchmarks for core skills so that regressions in quick, fix-tests, or debug are caught before merge.

Proposed Solution

Build a generic bin/nf-benchmark.cjs --skill=<name> --track=smoke|full --json runner that loads per-skill fixtures from benchmarks/<skill>/ and integrates into the existing benchmark-gate.yml alongside solve.

MVP scope: extract shared utils → generic runner → quick smoke fixtures → CI gate update.

Acceptance Criteria

• bin/nf-benchmark.cjs --skill=quick --track=smoke --json exits 0 and outputs valid JSON with pass_rate
• Smoke fixtures for nf:quick exist under benchmarks/quick/fixtures.json with at least one exits_zero fixture
• benchmarks/quick/baseline.json exists and the CI gate enforces it
• benchmark-gate.yml runs both solve and quick and fails if either score drops below its baseline
• Shared utilities (evaluatePassCondition, extractResidual, snapshot/restore) are extracted into bin/benchmark-utils.cjs and imported by both runners
• No fixture in the CI (smoke) track requires an LLM API key to execute

Decomposition

1. Extract evaluatePassCondition, extractResidual, snapshotFormalJson, restoreFormalJson from nf-benchmark-solve.cjs → bin/benchmark-utils.cjs
2. Build bin/nf-benchmark.cjs accepting --skill=, --track=smoke|full, --json, loading from benchmarks/<skill>/
3. Add benchmarks/quick/fixtures.json with smoke fixtures (exits_zero on --dry-run or --plan-only invocations)
4. Add benchmarks/quick/baseline.json with initial pass_rate floor
5. Update benchmark-gate.yml to run both solve and quick and check both baselines
6. Update nf-benchmark-solve.cjs to delegate to benchmark-utils.cjs (remove duplication)

Not Doing (this issue)

• LLM-calling fixtures in CI
• Fixtures for new-milestone, fix-tests, debug — add once the generic runner is proven
• Composite cross-skill scoring

Difficulty

M — Pattern is proven (solve runner exists), extraction + generalization is mechanical. Main risk: quick may not have a cheap deterministic smoke path, which could require fixture design work.

Open Questions

• Does nf:quick have a --dry-run or --plan-only flag that makes smoke fixtures cheap without LLM calls?
• Should baselines auto-update after merge or remain manual (currently manual for solve)?

[jobordu]

All acceptance criteria implemented and verified: generic runner, quick smoke fixtures, shared benchmark-utils.cjs, and CI gate updated.